基于时空上下文分层卷积的相关滤波算法

Correlation filtering algorithm based on spatio-temporal context hierarchical convolution

目标在跟踪过程中由于运动速度过快、跟踪精度不够、出现遮挡而经常会导致跟踪失败。针对这一问题,文中在目标跟踪算法KCF的框架下提出一种基于时空上下文分层卷积的相关滤波算法。其中选用训练好的VGG-19网络提取多层深度特征送入岭回归中训练模型,并对多层深度特征的响应值采用自适应分配权重的方式得出最终的置信度,同时依据上一帧预测位置选取与目标同等宽高的上下左右四块区域,增加在目标搜索区域的四周用来约束目标因速度过快超出搜索范围的情况,然后针对目标遮挡提出一种新型遮挡判断机制用以判断目标是否被遮挡,并配合峰值分布的方差更新模型,避免使用目标被遮挡所生成的污染样本来更新模型。经实验验证所提方法在目标快速运动和遮挡时能有效跟踪。

Due to excessively fast movement,insufficient tracking accuracy and occlusion,tracking failure always occurs in the process of tracking.In view of this,a correlation filtering algorithm based on spatio-temporal context hierarchical convolution is proposed under the framework of the target tracking algorithm KCF(kernelized correlation filter).The trained VGG-19 network is selected to extract the multi-layer depth features,and then the features are sent to the ridge regression to train the model.The response value of the multi-layer depth features is subjected to adaptively assigning weights to obtain the final confidence.At the same time,according to the predicted position of the previous frame,four areas(up,down,left and right)with the same width and height as the target are selected and added around the target search area to constrain the target from exceeding the search range caused by excessive speed.And then,a new occlusion judgment mechanism is proposed on account of target occlusion,so as to determine whether the target is occluded.In addition,the model is updated with the variance of peak distribution,while not updated by the polluted samples generated by the targets being occluded.The experiments show that the proposed algorithm can effectively track the target when the target moves fast or is occluded.